Commit Graph
2 Commits
Author SHA1 Message Date
host 99d9e5771b merge(chv-hybrid): 合入 B 轨——租约 + 排队 + 升级提示 + 网关门
合入 feat-chv-hybrid-b (667a427)。executor.py 两处内容冲突按 hunk
解决,全部取 B 侧;合入时发现 merge 复制出第二个 suspend_after_task
定义(A 版截断体),已删除。

集成统一(A/B 语义冲突三处,设计 §4.1 的 state 模型为准):
- state 模型:A 的 flag 模型(resident: True + legacy state)与 B
  的 state 模型(state ∈ resident-active/paused/hibernated/demoted)
  统一为 state 模型;_mark_resident_record 标记常驻时同时置
  resident-active,_ensure_resident_capacity 同时认 flag 与 state。
- 容量等待:A 的 bool 版 _wait_for_capacity 改名 _poll_capacity,
  B 的 raise 版保留原名;_wait_for_resident_slot 豁免已有常驻登记
  的租户(它已占一个槽位,再等等于把自己的槽位算成满——A 轨
  test_resident_cap_blocks_new_resident 第二腿在合入时因此失败,
  已修并通过)。
- 档位大小写:A 用小写 paid/heavy,B 用 Paid/Heavy;比较一律走
  _norm_tier + HYBRID_RESIDENT_TIERS(resident_tier_eligible 与
  promote_to_resident 两处),fail-closed 语义不变。
- 严格快照校验翻为默认开启(设计 §2.2):_snapshot_strict_tenant
  默认 True,显式 CHV_SNAPSHOT_STRICT_TENANT=0 才关闭;旧恢复
  测试的 fixture 已登记带匹配 tenant 的快照,A 轨 legacy 测试改
  为显式 opt-out 覆盖。

生产接线:worker 只调 dispatch_task,dispatch_task_hybrid 在生产
路径上无调用方;dispatch_task 顶部加一道门,warm_pool_size>=1 时
走 dispatch_task_hybrid,==0 时保持旧冷启语义(不成环:hybrid 的
<=0 分支退化为 dispatch_task_persistent,后者 <=0 调回 dispatch_task)。

旧网关测试更新(B 的语义变更):test_worker_gateway_guard_allows_chv
前加常驻登记;新增 test_worker_gateway_guard_denies_pool_vm。

测试:test_chv_hybrid_a/b 64 项 + 受影响旧测试 16 项(Gateway、
RestoreNoFallback、RestoreSnapshotTest)全部通过;全量待跑。
2026-11-08 22:00:12 +00:00
host 172accb2ec feat(chv): hybrid executor core — routing, clean pool, data disk, snapshot tenant check
Track A (§9.1 items 1-4, DESIGN-HYBRID.md):
- dispatch_task_hybrid: route by tier + lease state (resident resume/cold
  vs pool hit/cold/queue, 300s queue timeout -> host_capacity_exceeded)
- clean pool prewarm split from prewarm_persistent_vm: tenant-less birth
  (no data disk, no seed, no Gitea sync wait), binding deferred to
  acquire; warm_pool_sync_incomplete fail-closed at allocation
- tenant data disk mount at acquire (attach_tenant_data_disk, PUT
  /vm.add-disk + chv-data-mount) and unmount+sync at release;
  _verify_task_disk_source invariant: task disk must be the tenant's
  data disk, never the clean snapshot disk
- restore_vm validates tenant_id before restore: unregistered/unmarked
  snapshot denied in strict (hybrid) mode, mismatch -> snapshot_tenant_
  mismatch fail-closed, audit vm_restore_denied (host_id + tenant only)
- settled values: 2 warm per host, pool cap 50% of low-freq peak,
  resident cap 50% of VM slots, 1 snapshot per tenant per 30 days

Tests: cloud/tests/test_chv_hybrid_a.py (27 tests, all passing)
Full cloud suite: 4606 tests, 0 failures, 0 errors, 15 skipped
(TMPDIR on /home; ~91 min, slowed by concurrent track-C suite)
2026-11-08 21:45:03 +00:00